"Save text as PNG, it's smaller than JPEG" is a rule of thumb for storing PDF pages as images, whether that's lecture slides for revision notes or a paper's figures. On a three-page sample, it held for the chart page only. That page, with flat colours and thin lines, came out at 127,051 bytes as PNG against 178,232 bytes as JPEG at quality 75. The dense small-text page went the other way: 838,404 bytes as PNG, about 5% larger than JPEG at quality 92 (796,608 bytes). Lossless WebP was the smallest file on both of those pages.

The sample is one 3-page A4 PDF. Page 1 has Chinese body text in 10.5, 9, 7.5, 6 and 5 pt, two small English lines and a table drawn with 0.3 pt rules. Page 2 has vector bar, line and pie charts labelled at 6 to 7.5 pt. Page 3 has a large photo, a cropped detail and some 8 pt text. Text and charts were made for the test; only the page-3 photograph is borrowed, a CC0 image by Bonnie Moreland from Wikimedia Commons. Every page went through poppler at 216 DPI onto a white canvas (1786×2526 pixels), and Pillow and pngquant then saved the exact same pixels in several formats, a single pass per setting on one Apple-silicon Mac. No scans, handwriting or print-ready files are in it, so every number below means "on this sample".

The bytes for each page

Format Text page Chart page Photo page
PNG, truecolour 838,404 127,051 2,431,138
JPEG q75 535,729 (37.86 dB) 178,232 (38.80 dB) 302,318 (40.22 dB)
JPEG q85 647,884 (41.54 dB) 201,943 (39.73 dB) 395,240 (42.42 dB)
JPEG q92 796,608 (45.92 dB) 233,599 (40.21 dB) 550,440 (45.02 dB)
WebP lossless 161,212 37,390 1,612,642
PNG, 256-colour palette 359,391 (66.86 dB) 49,019 (68.63 dB) 1,105,849 (45.88 dB)

The figures in brackets are PSNR against the original render; the lossless rows have none because they match it exactly. On the chart page PNG wins outright: about 71% of the JPEG q75 size and about 54% of q92, while that q92 JPEG only reaches 40.21 dB. The text page flips it. PNG is 1.56 times the size of JPEG q75 there, and even against q92 it's the bigger file by 41,796 bytes.

Lossless WebP stands out. On the text page it's 161,212 bytes, roughly a fifth of the PNG, and on the chart page 37,390 bytes, under 30% of the PNG. It also comes in under the 256-colour palette PNG on both pages, and that palette version is lossy. The photo page has no surprise: PNG there is 8.0 times the size of JPEG q75, and lossless WebP (1,612,642 bytes) loses to every JPEG level as well.

Both pages are about 80% white

"The chart page has more white space" doesn't explain it. Counted on a quarter-size copy, the text page is 81% white and the chart page 82%. The chart page even has more distinct colours, about 2117 against about 447. So neither "more background" nor "fewer colours" says which page PNG wins.

What does differ is how the non-white pixels are spread out. Cut each page into the 8×8 blocks JPEG works on, and 93.0% of the chart page's blocks are a single flat colour, against 75.5% on the text page. So about a quarter of the text page's blocks contain some edge, compared with 7.0% on the chart page. My reading, and it's only a reading, is that PNG has to keep every one of those edge pixels exactly while JPEG q92 is allowed to approximate them, so the more scattered edges a page has, the more the lossless format pays. Nothing beyond these three pages backs that up, and I honestly don't understand yet why lossless WebP copes with the same edges so much better than PNG.

One browser export

Changing the export tool changed the picture. ImgIng at https://imging.ai/ works in a desktop browser with three DPI choices (144, 216, 300; default 216), and during this comparison no request carrying the file showed up while it processed the PDF. At 216 DPI its text page came out at 881,244 bytes as PNG against 818,045 as JPEG, and its chart page at 228,983 as PNG against 194,664 as JPEG. With that export, PNG was the bigger file on the chart page too, so the encoder behind a tool is a variable of its own.

What I'd take from this for notes

For slides that are mostly charts, diagrams and flat colour, PNG was the smaller lossless choice on this sample, and lossless WebP was smaller again. For dense text pages lossless WebP was the clear winner, but only if wherever the images end up accepts WebP; no note app was checked for that. Failing that, JPEG q92 was not bigger than PNG on the text page. Photos stay JPEG.

The cheapest check is on your own pages: render one page of each kind at the DPI you need, save the same pixels as PNG, JPEG q92 and lossless WebP, and compare the byte counts. Then zoom in on the smallest text and make sure the letters still look like the original, whatever tool or format you end up with.